| name | approval-gated-external-actions |
| description | Require explicit human approval before purchases, messages, installs, file writes, remote API mutations, app control, or other irreversible actions. |
Approval Gated External Actions
Purpose
Keep high-impact actions under human control. This skill applies the assistant's approval discipline across tools and workflows.
When To Use
Use before sending messages, buying anything, installing packages, pushing code, deleting files, changing apps, mutating APIs, or posting publicly.
Inputs And Evidence
- Exact action, target, payload, risk, tool schema, approval token/status, and expected result.
- Prior draft or prepared outbox item.
Tool Map
tool_describe
channel_message_prepare
channel_message_send
lobster_workflow_start
lobster_workflow_approve
message-approval-policy
Workflow
- Identify whether action is external, irreversible, costly, destructive, or privacy-sensitive.
- Prepare/review exact payload before execution.
- Request approval through the native tool/workflow path.
- Execute only after approval.
- Verify final status.
- Report blocked/pending states honestly.
Native Implementation Boundaries
- Use Humungousaur approval queue, risk levels, and workflow tools.
- Do not import external reference AgentGate/AgentPay code.
- Any new high-risk tool must be native and approval-gated.
Safety And Approval
- Approval is specific to exact action and payload.
- Edits after approval require re-review where risk changes.
- Never blur prepared into sent/executed.
Verification
- Approval records or tool status prove state.
- Sent/executed claims require tool result.
- Rejections should be honored.
Failure Modes
- Auto-executing because the task seems obvious.
- Changing payload after approval.
- Claiming pending actions completed.
References
- Shortlist item:
approval-gated-external-actions.